Information About the Project

The position is announced at Chalmers University of Technology.

Learning and Leveraging Rich Priors for Factorization Problems
In this project we are interested in developing methods that combine traditional (parametric) mathematical formulations induced by domain expertise with (non-parametric) models learned from examples. Parametric models inject domain knowledge into learning-based approaches and have therefore the potential to massively reduce the necessary amount of training data. Additionally, the output can be constrained e.g. to be physically plausible, which is difficult to guarantee with pure learning-based architectures. At the same time, being able to incorporate learned priors has the potential to regularize problems where a physical model is not sufficient to guarantee a well posed formulation. From a theoretical point of view we are interested in results that characterize formulations in terms of their expressiveness and generalization as well as developing efficient inference approaches.

Our main application of interest are factorization-based problems, in particular non-rigid structure-from-motion (NRSfM), which aims to infer 3D models of dynamic scenes or objects from videos or multiple images. In contrast to its rigid counterpart, NRSfM is far less mature and it is inherently an ill posed problem requiring suitable priors that disambiguate the effects of camera motion and object deformation. Hence, this project enables research at the intersection of 3D computer vision, factorization methods and machine learning.